HEEF: Hierarchical Event Extraction via Dual-Stage Semantic Compression and Chain-of-Thought Reasoning

Event extraction (EE) is a fundamental task in natural language processing. However, it remains challenging in complex document scenarios. On one hand, the intricate contextual dependencies and long-text structures are difficult to model accurately. On the other hand, large language models (LLMs) for EE tasks face substantial computational resource demands and task-specific adaptability challenges. To address these challenges, we propose a Hierarchical Event Extraction Framework (HEEF). This framework leverages multi-level semantic compression and a chain-of-thought-driven entity generation paradigm to reduce computational resource consumption while maintaining entity extraction accuracy. Meanwhile, the hierarchical design of HEEF effectively captures long-range dependencies and mitigates event type confusion in complex scenarios. Experimental results on two benchmark datasets, ACE2005 and MAVEN-Arg, demonstrate that HEEF significantly outperforms existing methods, achieving F1 score improvements of 5.1% and 7.7% over the best baseline, respectively, validating the effectiveness of the HEEF framework.